{"id":"11fc3e4a-be20-4d42-8c37-bc61a707a56f","arxiv_id":"2411.08216","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GTA, a post-processing module, splits and merges player tracklets using ReID features and clustering, improving HOTA scores on SportsMOT and SoccerNet across three trackers.","lead":"This paper adds a post-processing step that cleans up player tracking by splitting mixed tracklets and reconnecting fragments of the same player using appearance features. It reports higher tracking accuracy on SportsMOT and SoccerNet across three trackers, and the code is released online.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing risk is ReID training/evaluation overlap: Section 4.2 says OSNet is 'trained on SportsMOT' without stating the train/test split, so the reported gains could be inflated by identity leakage from test videos.","rationale":"I looked for the most load-bearing threat to the central claim that GTA is a plug-and-play post-processor giving consistent, large gains across trackers and datasets. The reader's identified weakest assumption matches mine: the ReID model is the only learned component, and the paper's one-sentence description of its training data creates a concrete leakage risk that would invalidate all reported improvements if realized. I considered other risks: fixed hyperparameters could overfit a particular dataset, the spatial constraint in Eq. (4) is loose when beta=1, and the absence of comparisons with TransLink, AFLink, and similar global-link methods weakens the SOTA claim. None of these is as damaging as possible test-set contamination of the ReID embeddings, because those embeddings feed both the splitter and the connector and are the sole source of appearance information. A single check, retraining OSNet on the official SportsMOT train split and rerunning GTA, would settle the matter. Since this is a clarification-and-verification concern rather than evidence of a false result, the existing conditional verdict is appropriate.","tokens_in":8875,"tokens_out":5642,"duration_ms":59700,"concrete_test":"Check the released repository and checkpoint metadata to identify the exact OSNet training split (e.g., the training image or sequence list). Then retrain OSNet from scratch on the official SportsMOT train split only, keep all GTA hyperparameters fixed (s=5, epsilon=0.6, k=3, alpha=0.4, beta as in Section 4.2), and re-run GTA on Deep-EIoU, ByteTrack, and SORT over the SportsMOT test set. If HOTA or IDF1 gains on Deep-EIoU drop by more than roughly 1 point, or if the relative ordering of improvements changes, the leakage concern is confirmed; if the gains are reproduced, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 states, 'For our experiments, we use the OSNet model trained on SportsMOT dataset,' and gives no further detail about the ReID training protocol. SportsMOT has an official train/test split; if the OSNet checkpoint was trained on the full dataset or on sequences overlapping the SportsMOT test videos, the appearance embeddings used by both the Tracklet Splitter and Tracklet Connector would encode the actual test identities. Since GTA's only learned component is this ReID feature extractor, such leakage would directly inflate the association metrics (AssA, IDF1, HOTA) and the reported ID-switch reductions in Tables 1 and 2, and it would undermine the plug-and-play claim for unseen sports footage. Even if the authors intended 'trained on SportsMOT' to mean the official training split, the paper does not say so, so the released code and checkpoint must be checked. The missing comparison with prior global-link methods is secondary: it affects the SOTA label but not the internal consistency of the reported improvements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GTA (Global Tracklet Association), a two-stage post-processing method for multi-object tracking in sports. A Tracklet Splitter uses DBSCAN clustering on OSNet ReID embeddings to split tracklets that contain multiple identities, and a Tracklet Connector uses hierarchical clustering with cosine-distance and spatial constraints to merge fragmented tracklets of the same identity. The method is evaluated by applying it to SORT, ByteTrack, and Deep-EIoU on SportsMOT and SoccerNet, reporting large improvements in HOTA, AssA, IDF1, and reductions in ID switches, including HOTA of 81.04% on SportsMOT and 83.11% on SoccerNet. The code is open-sourced.","tokens_in":9157,"tokens_out":3483,"duration_ms":34374,"significance":"If the reported results hold, GTA would be a simple, reusable, post-hoc refinement tool with consistent benefits across multiple trackers and two sports datasets, which is practically valuable. The paper's strengths include the clear two-stage design, the consistency of the improvements across SORT, ByteTrack, and Deep-EIoU, and the open-source release. The main uncertainties are the provenance of the ReID model used for all appearance comparisons and the lack of sensitivity analysis for the five hyperparameters, which together affect the verifiability and generalizability of the central claims. The state-of-the-art claim also depends on comparisons with existing global-link methods that are currently missing.","major_comments":[{"comment":"The statement 'we use the OSNet model trained on SportsMOT dataset' does not specify whether this means the official SportsMOT training split. Since Table 1 reports results on the SportsMOT test set, a checkpoint trained on the full dataset or on sequences overlapping the test videos could encode test-time identities, directly inflating AssA, IDF1, HOTA, and the reported ID-switch reductions in Tables 1 and 2. The authors must state the exact split used for ReID training and confirm, for example by releasing the checkpoint and training script, that no SportsMOT test sequence was seen during ReID training; otherwise the central plug-and-play claim and the magnitude of the improvements are not verifiable.","section":"§4.2, ReID Model"},{"comment":"Five free parameters (epsilon, s, k, alpha, beta) are fixed without sensitivity analysis, and beta is set differently per dataset (1 for SportsMOT, 0.7 for SoccerNet). Given that no error bars or per-sequence variance are reported, it is unclear how robust the gains are to reasonable perturbations of these choices, and whether the per-dataset beta is a principled setting or a form of test-set tuning. I request a sensitivity table varying each parameter around its chosen value, or at least results with a shared beta across both datasets.","section":"§4.2, Hyperparameters"},{"comment":"Prior global tracklet association methods (Translink, AFLink, Split and Connect, MambaTrack) are described in Related Work but are not compared against in the experiments. Because the abstract claims state-of-the-art performance, the absence of these baselines leaves the SOTA claim unsupported. Please add comparisons with at least one or two existing global-link methods under a shared ReID setting, or weaken the claim to describe improvements over the selected online trackers.","section":"§2.3 and Tables 1-2"}],"minor_comments":[{"comment":"The text says Deep-EIoU achieves 85.4% HOTA on the SoccerNet test set published in 2022, while Table 2 reports the Deep-EIoU baseline as 79.41%; please clarify which SoccerNet split/version is being used and why the numbers differ.","section":"§4.2, Tracker"},{"comment":"The notation reuses i and j for both tracklet indices and frame indices, and the summation is written over 'i∈Πi' and 'j∈Πj'; please use distinct frame indices (for example p and q) and clarify the temporal-overlap condition in the first case.","section":"§3.2, Eq. (1)"},{"comment":"The spatial constraint is defined for 'beginning and ending' points of tracklets, but the text does not state whether the constraint is applied symmetrically or how it interacts with tracklets whose temporal spans overlap; please make this explicit.","section":"§3.2, Eq. (4)"},{"comment":"The text says the method reduces ID switches and 'tracklet fragments (Frag)', but Frag is not reported in Tables 1 or 2; either include it or remove the mention.","section":"§4.3"},{"comment":"The baseline rows are not labelled with the method name in the leftmost column, which makes the table hard to parse; please add explicit method names to those rows.","section":"Table 3"},{"comment":"The abstract states 'consistently increasing the HOTA score from 79.41% to 83.11%' without identifying Deep-EIoU as the baseline; please name the baseline tracker and note that these are single-run results without error bars.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The ReID training/test overlap is the make-or-break issue for this paper. If the authors cannot confirm that the OSNet checkpoint was trained only on the official SportsMOT training split, the reported gains should be re-evaluated. Please also ensure that the released code and checkpoint correspond exactly to the experimental settings in the paper, including the per-dataset hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a clean, empirical tracklet post-processor for sports MOT. It splits mix-up tracklets with a DBSCAN variant and merges fragments hierarchically with spatial gates. That combination is modestly new relative to Split-and-Connect and global link work, and the gains are real: consistent HOTA/AssA/IDF1 jumps across three trackers and two datasets, with the biggest wins on the weakest tracker. The code is public and the ablations are sensible.\n\nThe strongest result is that the connector alone gives most of the gain; the splitter adds a smaller boost. The method is genuinely plug-and-play in the practical sense: it consumes tracklet boxes and a frozen ReID embedding.\n\nThe main soft spot is the ReID training description. Section 4.2 says only \"OSNet model trained on SportsMOT dataset\" with no mention of train/test split. Since the only learned component is exactly that embedding, if the checkpoint was trained on any evaluation video or on the full dataset, the association metrics are inflated and the plug-and-play claim for unseen footage collapses. This is concrete and checkable: the code and checkpoint are public, so an editor should ask the authors to state the split and, ideally, also report results with a ReID model trained only on the official training split or on held-out sports data.\n\nSecondary soft spots: no hyperparameter sensitivity (epsilon, s, k, alpha, beta are hand-set; beta changes by dataset), no error bars, and no comparison against existing global-link methods like Translink, AFLink, or MambaTrack. The missing comparisons matter for the SOTA label but not for the internal consistency of the reported improvements. The missing sensitivity analysis matters because the method has five free parameters and the paper does not show how stable the gains are.\n\nOverall, I believe the paper is honest and the improvements are very likely real, assuming the ReID checkpoint is clean. It is a solid workshop-to-journal contribution, not a breakthrough. It deserves serious peer review, and the referee should make the ReID split explicit and run the public code to verify the numbers.","headline":"A clean, empirical tracklet post-processor that likely delivers real gains; the ReID training split must be clarified before the SOTA claim is trusted.","tokens_in":9650,"tokens_out":1606,"would_cite":true,"duration_ms":17315,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A plug-and-play post-processing step that splits mixed tracklets and rejoins fragmented ones raises sports multi-object tracking accuracy on two benchmarks.","keywords":["multi-object tracking","tracklet association","tracklet refinement","re-identification","sports analytics","DBSCAN clustering","hierarchical clustering","HOTA"],"falsifier":"Train the OSNet ReID model only on the official SportsMOT training split, apply GTA to the test split, and compare with the paper's reported HOTA values; if the gains vanish, the results relied on test-set exposure. A second check is to run GTA on a fixed-camera video where two teammates wear near-identical jerseys, and count whether the splitter creates false splits and the connector merges distinct players.","tokens_in":8734,"feed_emoji":"⚽","tokens_out":6437,"duration_ms":56949,"temperature":0.7,"pith_summary":"The paper argues that the two dominant failure modes in sports player tracking are fixable after the fact by a global post-processing stage. Mix-up errors put several identities into one tracklet, while cut-off errors fragment one player across many tracklets. The proposed Global Tracklet Association (GTA) splits mixed tracklets by clustering per-box appearance embeddings with DBSCAN, then merges fragments that belong to the same player using hierarchical clustering under spatial and temporal constraints. Applied to three existing trackers, GTA raises HOTA on SportsMOT to 81.04% and on SoccerNet to 83.11%, improving SORT, ByteTrack, and Deep-EIoU across association metrics. If correct, it means tracker accuracy can be improved without retraining the tracker itself, only by cleaning its output.","feed_headline":"Tracklet split-and-merge lifts sports tracking HOTA to 81%","feed_subtitle":"A plug-in post-processing step raises HOTA on SportsMOT and SoccerNet for SORT, ByteTrack, and Deep-EIoU.","key_machinery":"The machinery is the Tracklet Splitter and Tracklet Connector pair. The splitter applies DBSCAN to per-box OSNet appearance embeddings, using minimum samples, a cosine-distance neighbor threshold, and a maximum-cluster parameter, and reassigns outliers to the nearest cluster instead of discarding them. The connector builds a pairwise cosine-distance matrix that forbids merging temporally overlapping tracklets, applies spatial thresholds derived from the maximum horizontal and vertical box-center distances scaled by a factor, and merges via hierarchical clustering up to a distance threshold. All split and merge decisions reduce to comparing OSNet appearance embeddings, so the ReID model is the shared foundation of both stages.","core_discovery":"GTA is a two-stage tracklet refinement pipeline that runs after a tracker has produced its raw output. In the splitter stage, each tracklet's box-grained appearance features, extracted by an OSNet ReID model, are clustered with a modified DBSCAN that assigns outliers to the nearest cluster and enforces a maximum number of clusters to avoid over-fragmentation. In the connector stage, a symmetric cosine-distance matrix over all tracklet pairs is built, with distance forced to 1 for temporally overlapping tracklets and for pairs whose exit/entry points violate spatial thresholds scaled to the field size, after which hierarchical clustering merges fragments until no pair distance exceeds a threshold. On SportsMOT the full pipeline improves Deep-EIoU HOTA from 77.21% to 81.04%, and on SoccerNet from 79.41% to 83.11%, with gains concentrated in AssA and IDF1 and reduced ID switches.","pith_inferences":["If ReID embeddings are the load-bearing component, then swapping OSNet for a stronger or domain-adapted ReID model should produce further gains; this is directly testable with the released code.","The same split-then-connect recipe could be applied to other tracklet-producing domains, such as vehicle tracking or multi-camera pedestrian tracking, wherever mix-up and cut-off errors dominate.","The spatial constraint assumes players re-enter from the same side of the field; a camera that pans would break that assumption, so a relaxed or learned spatial prior would be needed for such footage."],"forward_implications":["Any existing tracker can be improved by running GTA on its output; SORT, ByteTrack, and Deep-EIoU all gain HOTA, AssA, and IDF1 while losing ID switches.","Because GTA changes association rather than detection, it leaves DetA and MOTA nearly unchanged, so the improvements come almost entirely from fixing identity assignment.","The connector's spatial constraint encodes a fixed-camera, same-side re-entry assumption, so the method transfers directly to static sports broadcasts but would need adjustment for panning or moving cameras.","Applying GTA to Deep-EIoU yields state-of-the-art results on both benchmarks, with HOTA 81.04% on SportsMOT and 83.11% on SoccerNet."],"supporting_citations":[{"why":"Supplies the OSNet ReID model that produces the box-grained appearance embeddings used by both the splitter and the connector.","marker":"[32]"},{"why":"Provides the DBSCAN clustering algorithm that the splitter adapts to group tracklet boxes and reassign outliers.","marker":"[11]"},{"why":"Deep-EIoU is the strongest baseline tracker that GTA is applied to, and its SportsMOT and SoccerNet results are the ones extended to state-of-the-art.","marker":"[19]"},{"why":"Defines the SportsMOT benchmark on which GTA reports its 81.04% HOTA result.","marker":"[8]"},{"why":"Defines the SoccerNet-Tracking benchmark on which GTA reports its 83.11% HOTA result.","marker":"[7]"},{"why":"Supplies the HOTA metric that the paper uses to measure association-focused tracking performance.","marker":"[22]"},{"why":"Provides the hierarchical clustering approach that the connector uses for merging tracklet fragments.","marker":"[14]"},{"why":"Prior split-and-connect tracklet booster that addresses both mix-up and cut-off errors, positioning GTA's contribution.","marker":"[25]"}],"fun_headline_variants":["Tracklet split-and-merge lifts sports tracking HOTA to 81%","Global tracklet association boosts sports tracking HOTA to 81%","Plug-and-play tracklet refinement lifts sports tracking HOTA to 81%","One tracklet refinement step lifts HOTA to 81% in sports tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything depends on the ReID model's appearance embeddings reliably separating different players and recognizing the same player across long gaps; the paper states the OSNet model is trained on SportsMOT but does not specify its train/test split, so if those embeddings were exposed to test-video identities, the reported gains could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Tracklet split-and-merge lifts sports tracking HOTA to 81%","Global tracklet association boosts sports tracking HOTA to 81%","Plug-and-play tracklet refinement lifts sports tracking HOTA to 81%","One tracklet refinement step lifts HOTA to 81% in sports tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001165,"raw_usage":{"total_tokens":4818,"prompt_tokens":935,"completion_tokens":3883,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":3802}},"tokens_in":551,"tokens_out":3883,"duration_ms":25755,"temperature":1.0,"reasoning_tokens":3802,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:50:20.228266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the OSNet ReID model only on the official SportsMOT training split, apply GTA to the test split, and compare with the paper's reported HOTA values; if the gains vanish, the results relied on test-set exposure. A second check is to run GTA on a fixed-camera video where two teammates wear near-identical jerseys, and count whether the splitter creates false splits and the connector merges distinct players.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the OSNet ReID model that produces the box-grained appearance embeddings used by both the splitter and the connector."},{"cited_title":"In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96)","cited_arxiv_id":null,"evidence_quote":"Provides the DBSCAN clustering algorithm that the splitter adapts to group tracklet boxes and reassign outliers."},{"cited_title":"In: Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision","cited_arxiv_id":null,"evidence_quote":"Deep-EIoU is the strongest baseline tracker that GTA is applied to, and its SportsMOT and SoccerNet results are the ones extended to state-of-the-art."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Defines the SportsMOT benchmark on which GTA reports its 81.04% HOTA result."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Defines the SoccerNet-Tracking benchmark on which GTA reports its 83.11% HOTA result."},{"cited_title":"In: CVPR workshops","cited_arxiv_id":null,"evidence_quote":"Provides the hierarchical clustering approach that the connector uses for merging tracklet fragments."},{"cited_title":"IEEE Transactions on Multimedia25, 1256–1268 (2022)","cited_arxiv_id":null,"evidence_quote":"Prior split-and-connect tracklet booster that addresses both mix-up and cut-off errors, positioning GTA's contribution."}],"review_version":1}